Try removing
“-mtune=native -march=native” from the CFLAGS. It is likely that you have a heterogeneous server architecture on your cluster. The model is likely compiled once on one node, and then tried to run on a different node with a different CPU, and so it fails. If you remove those CFLAGS, it will be compiled to work on any cpu.
I had a similar issue and spent some time creating a script that would inspect the CPU architecture and maintain a separate copy of the compiled model for each installation. However, it turned out that the -march=native -mtune=native flags didn’t really speed things up very much, so I abandoned this strategy. I also found that the -O3 flags added a lot to compile time, and did not produce the speed improvements that I expected, so I now use:
CXXFLAGS= -O2 -pipe -Wno-unused -DEIGEN_NO_DEBUG -DBOOST_RESULT_OF_USE_TR1 -DBOOST_NO_DECLTYPE -DBOOST_DISABLE_ASSERTS -DFUSION_MAX_VECTOR_SIZE=12 -DNO_FPRINTF_OUTPUT -std=c++11